CARE-AI: Cancer assessment risk estimation using artificial intelligence.

Abstract

e13618 Background: Access to genetic testing for detecting hereditary cancer is limited, and conventional methods for selecting candidates are often imprecise. As a result, only a small proportion of those tested are found to carry pathogenic mutations, highlighting suboptimal selection criteria. Variability in guidelines, subjective interpretations, and incomplete family histories further complicate candidate identification. This study aims to address these limitations by developing an artificial intelligence (AI) tool to enhance the selection process for germline genetic testing. Methods: The dataset comprises over 10,000 pedigrees from the Familial Cancer Unit of Hospital Universitario Ramón y Cajal. Each record includes structured data on familial relationships and demographic attributes, as well as semi-structured clinical annotations. A robust preprocessing pipeline was developed to address heterogeneity in family structures, standardize free-text clinical annotations, and impute missing values using statistical and categorical imputation methods. Three machine learning models, Gradient-Boosted Trees (XGBoost), Random Forests, and Neural Networks, were evaluated for predictive performance using an 80/20 training-testing split. The final model was selected based on its ability to predict pathogenic or likely pathogenic genetic variants with high sensitivity and specificity. An interactive user interface was developed to enable real-time predictions and seamless integration into clinical workflows. Results: The XGBoost model achieves an area under the curve (AUC) of 0.72, outperforming conventional statistical models in both accuracy and generalization across diverse cancer syndromes. Traditional models often lack specificity for hereditary cancers and are limited in their applicability to multiple cancer types. Conclusions: This AI-powered tool offers a significant advancement over existing methods by leveraging family and clinical data to improve the identification of hereditary cancer risk. With its robust generalization across cancer types, the tool enables more accurate and equitable access to genetic testing. Its integration into family cancer clinics promises to enhance resource allocation and increase detection rates of pathogenic variants, ultimately improving patient outcomes.

Ähnliche Arbeiten